Akari

If you develop on Windows you have watched this whole category ship for macOS first and get to you later. This is the opposite: an installer, a portable build, and a desktop that treats Windows as the target rather than the afterthought. The deciding trait is simply that it is here and it works where you are.

You are the wrong buyer if you are on a Mac, because there is no build for you and the alternatives there are better funded. Pick it if Windows is where you work and you run more than one agent. Pick Arbor if you have a choice of machine.

Squash, merge and rebase are built into the review surface, which means history rewriting sits one click away from a diff an agent produced and a human skimmed. Nothing in the row describes a guard on which branch a strategy may touch, or a check that a session's work was read before it was collapsed into a single commit.

It does one thing better than larger tools. Sessions survive a reconnect with their output intact, so a dropped window does not cost you the record of what an agent did while you were away.

1. The design normalises unlike things behind one interface. Several coding CLIs and a bare shell are addressed through adapters, which is the correct place to absorb their differences, though it also means the abstraction is only as good as its worst adapter. 2. Status, output and diffs propagate over a socket rather than by polling.

2. No evaluation is published, which is consistent with a product claiming coordination rather than capability. The isolation unit is a working tree per session, a choice this project borrowed rather than invented, and borrowing it was correct.

233 stars and one author. That is not a company, it is a personal tool published generously, and the longevity question reduces to whether that person still has this problem in a year. There is no entity to acquire, no revenue to defend, and no pricing power because there is no price.

Moat: none, and the ideas here are visible to anyone reading the repository. Likely path: quiet maintenance, then a slow stop, with the concepts surviving inside somebody else's project. Position: use it if it fits today, and keep the fork you would need if it stops.

Zero licence cost and a desktop application my endpoint team already knows how to manage. The governance side is empty: no SSO, no audit trail, no retention policy and no console, so what sixty engineers do with it is visible only in the repositories afterwards.

It does not run unattended, so it contributes nothing to a pipeline and cannot be measured as one. The seat cost is zero and the inference cost is not, and nothing here aggregates the second number. Approved with conditions: standard desktop management, and keys issued centrally.

MIT, which means if the author stops I can keep going. What I care about here is the shell: multiplexing defaults to PowerShell 7, so my profile, my modules and my aliases are what the agent works inside, rather than a stripped environment that behaves differently from my own.

There is no MCP here, in either direction, so nothing I have built as a server plugs in and I am back to whatever tools the underlying agent carries. Building from source is pnpm install and pnpm dev:all, which is the whole ceremony, and I appreciate that.